Where Should I Draw the Line When Using AI?

Why this question will not leave me alone

AI image tools have changed how I work as a web developer. A mock-up that used to need a photographer, a stock subscription and a week of waiting can now exist before lunch.

But the same speed made me stop and ask a question I could not answer. Where did these pictures come from? This post is my attempt to answer it honestly, with the engineer's view of how the tools work and the plain view of someone who has to answer to clients.

Conflicted, and I think that is fair

My feelings are conflicted. As a web developer, I’m amazed by AI's speed especially in image generation to create website mock-ups. Its like the song generated in the CBS Mornings (2025) video. But I'm also alarmed by the unethical foundation AI brings up. The CNA Insider (2025) documentary, showing how artists' work is "stripped away" and "scraped" for training, feels like theft. It’s a "nightmare" (Venus Theory, 2024) where creators and artists are exploited.

Where the speed comes from

The speed is not magic. Many popular image generators are diffusion models. Midjourney does not publish its internals, so I cannot say exactly how it works, but the idea is the same. They start from random noise and remove it step by step until a picture matches your text prompt. To learn what "a smiling customer support agent" looks like, they are trained on enormous collections of images paired with captions, and for several well-known models much of that data was collected from the open web, as the CNA Insider (2025) documentary describes.

That is the part that feels like theft to many artists, and it is not only a feeling. Researchers have shown that these models can memorize individual training images and reproduce them. In one study, the team pulled more than a thousand training examples out of state-of-the-art models, including photographs of individual people and company logos (Carlini et al., 2023). So when I say I do not know where a face came from, that is a technical fact, not a figure of speech.

My advocacy, and the question that follows

This connects to my advocacy: "We must learn to harness the power of AI responsibly by using our wisdom to guide its power."

So what can I personally do?

When it got personal

I just lived this. I used Midjourney, a lot, for the storytelling and image elements in a website mockup, which was fast and creative. But when I was generating images of people, Saudi customer support agents. I felt a sudden, real alarm.

I realized I had no idea if that AI-generated face was "scraped" (CNA Insider, 2025) from a real person. What if it generated a likeness of a public figure, a prince, or just a police? In Saudi Arabia, that could violate the Anti-Cybercrime Law, carrying massive penalties of 500,000 SAR (equivalent to PHP 7,881,000) (Nawaf Law Firm, 2025).

Why a generated face is a real risk

Most generated faces are blends of patterns. But the memorization research above shows a model can sometimes reproduce a real person's photo (Carlini et al., 2023), and that gap is the problem. The tool will not tell me whether a result happens to look like a real, identifiable person, and a client in a country with strict privacy rules will not accept "the model made it up" as an answer.

The Saudi Anti-Cyber Crime Law, as it is summarized in public legal guides, punishes invasion of privacy and defamation through technology with up to one year in prison and a fine of up to 500,000 riyals (Nawaf Law Firm, 2025; Salama, 2021). I am a developer, not a lawyer, and I am not saying a generated image breaks that law. I am saying the exposure is large enough that guessing is not a plan.

The wider legal fight is also still open. A group of artists sued Stability AI, Midjourney and DeviantArt in a class action over the use of their work as training material (Andersen v. Stability AI Ltd., 2023). In June 2025, Disney and Universal sued Midjourney, accusing it of copyright infringement (Disney Enterprises, Inc. v. Midjourney, Inc., 2025). Courts have not settled these questions, and I do not want a client's launch to be the test case.

My personal action

My personal action, therefore, is conscious limitation. While I use Midjourney for abstract concepts, my "wisdom" now forces me to draw a hard line. I will not use tools like Midjourney to generate a human likeness for a client. The ethical and legal risk is too high.

The rule, written as code

I like rules I can check, so here is mine as a small function. It is a simplification, but it is how I now think about every asset that goes into a client mock-up.

type Asset = {
  source: "ai" | "licensed" | "client-supplied";
  showsRealPerson: boolean; // a face, a body, anyone identifiable
  hasConsent: boolean;      // model release or written permission
};

function canShipToClient(asset: Asset): boolean {
  // The hard line: no AI-generated human likeness, ever.
  if (asset.source === "ai" && asset.showsRealPerson) return false;

  // Real people need real consent.
  if (asset.showsRealPerson && !asset.hasConsent) return false;

  return true; // abstract art, textures, backgrounds, icons
}

In practice that means:

• AI for abstract concepts, textures, gradients and backgrounds.
• Licensed stock with a model release, or the client's own photos, for any person on screen.
• Illustrations or simple placeholders when there is nothing else.
• A written note in the project file about which tool made which asset.

What I hope happens next

I hope for the 'Fairly Trained' (Newton-Rex, 2025) certification solution to thrive and be the enforced global standards of the popular generative AI companies. For now, my personal action is to set critical boundaries on AI tools starting from myself.

Certification is one piece. Another is provenance. The C2PA standard (Coalition for Content Provenance and Authenticity, n.d.) attaches signed Content Credentials to an image that record who made it, with which tool, and whether AI was involved. It does not fix the training data problem, but it makes it harder to pass off a generated face as a photograph.

Closing thoughts

I am not going to stop using AI. It is too useful, and pretending otherwise would be dishonest. But the speed is a reason to slow down at exactly one place: anywhere a real person might be hiding in the result. Draw your own line before a client or a court draws it for you.

References

• Andersen v. Stability AI Ltd. (N.D. Cal. 2023). Class action filed by visual artists against Stability AI, Midjourney and DeviantArt.
• Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramèr, F., Balle, B., Ippolito, D., & Wallace, E. (2023). Extracting training data from diffusion models. USENIX Security Symposium. https://arxiv.org/abs/2301.13188
• CBS Mornings. (2025, August 16). AI-generated music sparks industry concern [Video]. YouTube.
• CNA Insider. (2025, April 12). Copy + paste + steal: Artists battle for copyright vs generative AI | Undercover Asia | Full episode [Video]. YouTube.
• Coalition for Content Provenance and Authenticity. (n.d.). Content Credentials and the C2PA specification. https://c2pa.org
• Disney Enterprises, Inc. v. Midjourney, Inc. (C.D. Cal. 2025). Complaint filed June 11, 2025.
• Nawaf Law Firm. (2025, September 16). Fines for filming and defamation in Saudi Arabia according to the laws and penalties. https://nawaf-law.com.sa/en/غرامة-التصوير-والتشهير
• Salama, S. (2021, April 20). Saudi Arabia: Jail time, SR500,000 fine for filming people. Gulf News. https://gulfnews.com/world/gulf/saudi/saudi-arabia-jail-time-sr500000-fine-for-filming-people-1.78618999
• TED. (2025, March 19). How AI models steal creative work, and what to do about it | Ed Newton-Rex | TED [Video]. YouTube.
• Venus Theory. (2024, November 12). AI copyright claimed my last video [Video]. YouTube.